Unsupervised feature selection via row-sparse local preserving projection.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Unsupervised dimensionality reduction is crucial for high-dimensional unlabeled data.
• Local Preserving Projection (LPP) is a feature extraction method, while feature selection is also needed.
• Existing LPP-based methods use approximations (ℓ_2,p-norm) for feature selection, leading to suboptimal results.

Purpose of the Study:

• To propose a novel unsupervised feature selection method, Unsupervised Feature Selection via Row-Sparse Local Preserving Projection (UFSLP).
• To directly address the ℓ_2,0-norm constraint for optimal feature subset selection.
• To improve clustering accuracy and normalized mutual information in high-dimensional data.

Main Methods:

• Developed UFSLP, an unsupervised feature selection method.
• Preserves local neighborhood structure during feature selection.
• Incorporates Principal Component Analysis (PCA) as a regularization term to balance local and global information.
• Reformulates and solves the ℓ_2,0-norm optimization problem using a coordinate descent method.

Main Results:

• UFSLP effectively performs unsupervised feature selection.
• The method balances local and global information.
• Experiments show UFSLP outperforms state-of-the-art unsupervised feature selection methods on nine benchmark datasets.
• Achieved superior clustering accuracy and normalized mutual information.

Conclusions:

• UFSLP offers a robust solution for unsupervised feature selection.
• Directly optimizing the ℓ_2,0-norm leads to optimal feature subsets.
• The proposed method demonstrates significant improvements in data clustering tasks.